I think that in many of these cases the deciding factor is not only fluency in the language, but also the harder problem of context. Labelling individual sentences without context is hard enough, but what makes it worse is that it then spreads to the mistaken assumption that sentences can be analysed without context based on the initial training.

I would argue that the very idea of "sentiment analysis" as applied to individual tweets is flawed... and that's even before we get into the much, much harder problems of sarcasm and irony.